Digital Machine Learning Circuit for Real-Time Stress Detection from Wearable ECG Sensor

This paper presents a digital machine learning circuit for classifying stress condition from chest ECG signal from a wearable sensor. To minimize hardware cost, we use only 5 time-domain features that have much lower area and power consumption than frequency domain or combination of time and frequency domain features as is used conventionally. We test the time-domain features on several machine learning algorithms. Random Forest classifier shows the best classification accuracy of 0.96 with the time-domain features at an estimated power consumption of only 1.16mW at 65nm CMOS process which demonstrates feasibility of embedding a machine learning classifier in a wearable ECG sensor for real-time, continuous stress detection.